The features, and what they actually do
Every feature is described with its limit. A tool that does not say where it stops is a tool you discover by getting it wrong, on a document already sent. Two families here: documents, which work on a plan of pages, and images, which are transformed one at a time.
One single work plan
The services on the market cut the same act into five pages: “merge”, “split”, “delete pages”, “rotate”, “reorder”. Five uploads, five downloads, five round trips to prepare one file. Yet they are five ways of saying the same thing: which pages, in what order.
Here you drop in whatever you have (PDFs, photos, scans), you arrange, you export. A merge is a plan where nothing was removed; a split is a plan where everything was removed but three pages. Every page is shown as a thumbnail: you do not delete “pages 4, 7 and 12” of a report without seeing them.
Images → PDF, and PDF → images
From images to a PDF
Three phone photos of a receipt that an office wants to receive as a single PDF: drop them in, they become pages. From there nothing tells them apart from a PDF: you reorder them, rotate them, slip them between two pages of a contract, number them.
Each image is placed in the middle of an A4 or Letter sheet, without distortion and without forced enlargement; a small image stays small rather than showing its pixels. The accepted formats are JPEG and PNG: the two a PDF can carry as they are. The EXIF data of your photos does not survive: only the pixels are copied over, not the camera, not the date, not the GPS coordinates.
From a PDF to images
Every page comes out as PNG or JPEG, at 72, 150 or 300 dots per inch. Beyond one page the files arrive in a ZIP archive: no browser lets forty downloads start in a row.
The images are rendered from the document you produced, with its order, its rotations, its numbering and its watermark, not from the original files. What comes out therefore looks exactly like what the screen was showing.
A page turned into an image is flattened: no more selectable text, no more form fields. Sometimes that is exactly what you want. It is not redaction for all that: what is drawn stays readable to the eye, and therefore to character-recognition software.
Numbering
You have just merged three PDFs, taken eight pages out and reordered twelve: the numbers printed on the originals mean nothing any more. Numbering therefore applies to the document you produce, from one end to the other, whatever the number of files it came from.
- The form: “3”, “- 3 -”, “3/12”, “Page 3 of 12”.
- The position: the nine points of the page, margin included.
- The first number, and the first numbered page: two separate settings, because a cover and three front pages are not numbered, and the pagination must still start at 1 on the fourth sheet.
- Recto-verso alternation: the numbers move to the outer edge, right on odd pages, left on even ones. Without it, half the numbers of a bound document fall into the fold.
Landscape pages are treated as such
On a table in landscape, the number is printed the right way up and in the right place.
Watermarking
“COPY”, “DRAFT”, “CONFIDENTIAL: do not circulate”. A watermark exists so that a page that has been photocopied, photographed or printed still carries the wording. It is therefore laid over the content, never under it, and in transparency, so that it can be seen without preventing reading.
Diagonally across the centre, or repeated as a tile over the whole page: the second is far harder to crop out. The size adjusts itself to each sheet, landscape included.
It is not protection. A watermark is a drawing on top of a page; a determined person removes it. It is a mention that follows the document as it changes hands, and saying so here saves you from believing otherwise elsewhere.
Adding text and images
A mention, a date, an exhibit number, a “RECEIVED ON” stamp, a scanned signature, a logo: placed where you want them, on every page, the first or the last, with their size, their tilt and their opacity.
That is the vast majority of real “edit a PDF” needs: filling in a printable form, initialling, dating, stamping, marking an exhibit with a number.
It is not a word processor, and it will not become one. We do not rewrite the content of a PDF, we do not replace a word inside a sentence, we do not reflow paragraphs: a PDF is a frozen layout, and the tools that claim to edit it guess, re-cut and damage. What is done here is what can be done honestly: adding a layer on top.
Text is written with the standard PDF fonts, which cover English, French and the languages of Western Europe. A character they do not know is reported while you type, never when you save.
Shrinking
A twenty-page contract in text weighs a hundred kilobytes; the same contract scanned weighs forty times more. What weighs in a PDF is the images: stored at full resolution because the photocopier scans at 600 dpi and no piece of software ever asked it not to. Shrinking is therefore resizing and re-encoding them, and nothing else.
- We never rasterise. The tools that announce “90% saved” on a text document turn it into images: the text is no longer selectable, nor searchable, and nobody was told. Here a page stays a page.
- We leave small images alone: logos, letterheads, signatures. The saving would be counted in bytes, the loss would be visible on screen.
- We leave whatever we cannot re-read with certainty: transparency, CMYK, JPEG 2000, fax. And the report shown after the export says how many images were left alone, and why.
Cleaning: what comes out, and what does not
The exported document is rebuilt, never patched: we start from an empty PDF and copy the wanted pages into it. That is not an implementation detail, it is the feature.
- What is deleted really is. A PDF modified “by addition” keeps its earlier versions inside the file: the page you think you removed is still there, under the new one. That is how redacted documents have been “un-redacted” after publication. Here, what is not copied does not exist.
- The metadata goes: author, title, producing software (which often carries the name of the workstation), and the dates, replaced by epoch zero rather than by the hour at which you happened to be working.
- Active content is defused: document JavaScript, actions on opening, attachments, XFA forms, annotations that launch a program. Ordinary links stay: they only fire on a click.
What is read at the door
Neither the file name nor its declared type is believed: both are written by whoever sends the file. It is the bytes that are read. And a password-protected PDF is refused rather than forced.
Cropping
A book scanned flat carries two centimetres of grey on three sides; a receipt photographed floats in the middle of an A4 sheet. You draw the frame to keep on the page, looking at it: not by filling in “left margin: 24 mm” and then opening the export to see where it landed. The frame is held as fractions of the format: a file made of four documents in A4, Letter and A3 is framed the same way throughout.
Cropping frames, it does not erase. A PDF page carries a crop box and readers only display what it delimits, but the content that sticks out stays in the file, and any editor can put the box back to its original size. That is true of every tool on the market; the difference is that we write it down. If what sticks out has to disappear for good, redacting is what you need.
Signing, and laying a layer over a document
Signing is the same gesture as adding an image: you drop the photo of your signature in as an ordinary file, you place it where you want it, you add the date and the wording. The produced document is rebuilt page by page: the signature is part of the page, it is not an annotation that peels off. The EXIF data of the photo does not survive: not the camera, not the date, not the GPS coordinates.
It is not an electronic signature within the meaning of the eIDAS regulation. Nothing here certifies your identity, timestamps the act or seals the document with a certificate: it is the equivalent of a photocopied handwritten signature, and it is worth exactly what the trust between the parties is worth. For an act that requires a qualified signature you need a trust service provider — and that provider, by construction, receives your document.
Overlaying lays one PDF over another: headed paper over a letter, a drawing frame over forty sheets, a stamp over an appendix. The layer adjusts to the format of each page without distortion, and its opacity can be set. It is laid on top, never underneath: sliding a background under the content would cost the page its selectable text, its links and its form fields.
Getting out what is inside the document
The text
All the text of a PDF in a .txt file, with a marker per page: “page 12” is what you write in an email, and you have to be able to find it again. Pages without text are named as such rather than left blank. On a scanned file, the first line says “none of the 12 pages carries extractable text”: that is an answer, not a breakdown.
The text is given as the document carries it: words cut at the end of a line stay cut, columns are not untangled, tables are not rebuilt. Guessing those things means being wrong in silence. There is no character recognition here, and there will not be: it would require a library of several megabytes fetched from somewhere else, which this site forbids itself.
The images
Not a photo of each page — that is [PDF to images](outil:pdf-vers-images) — but the embedded images: the three photos of the accident report, the scan of the receipt, the chart placed in the report. A JPEG comes out byte for byte: this is not a conversion, it is an extraction, and no generation loss is added to the one the scanner already applied. An image placed on thirty pages comes out only once.
Whatever cannot be re-read with certainty is left in the document: JPEG 2000, CCITT fax, CMYK, indexed palette — and the screen says how many images, and why. An image that comes out with the wrong colours is discovered by the recipient; a missing image that was reported is not.
Comparing two versions
Two versions of a contract, of an offer, of a brief: the tool says which words changed, page by page. It is the one where uploading costs the most: sending a document to a third party means entrusting it with one exhibit; sending both versions means entrusting it with the negotiation itself.
Pages are first matched by their resemblance, never by their rank: without that, a cover page added at the front would make the whole rest of the document look rewritten. The words of matched pages are then compared so as to show the smallest set that explains the move from one text to the other: a comma added stays a comma added, not a paragraph that blinks. Unchanged text is tightened around each change; what changed never is.
The report is saved as plain text, removed words prefixed with a “-” and added ones with a “+”: the convention of every comparison tool, the one you paste into an email without having to explain it.
This tool compares text, and nothing else: not images, not layout, not colours, not tables as tables. A scanned document carries no extractable text: two scans would therefore compare like two blank pages, and the screen says so rather than announcing two identical documents.
Images, the second family
The same gestures, on files handled ten times more often: shrinking a photo so it goes through by email, resizing it for a form that refuses anything above 2 MB, cropping it, converting it because a site only accepts JPEG. And it is exactly the same problem: the services that do this ask you to upload the photo — of an identity document, of a proof of address, of a child.
Six tools: [compress](outil:compresser-image), [resize](outil:redimensionner-image), [crop](outil:rogner-image), [convert](outil:convertir-image) between JPEG, PNG and WebP, [rotate](outil:pivoter-image) and [watermark](outil:filigrane-image). Everything goes through the browser's own image decoder, the very one that displays web pages: no dependency was added for this family, and so nothing new goes looking for anything at load time.
EXIF data does not survive. A phone photo carries the camera model, the date, the time and often the GPS coordinates of the place where it was taken. Only the pixels are copied over: what comes out of here no longer says where you were.
What we will not do: enlarge with artificial intelligence, cut out a background, retouch a face. Those three require a model of several tens of megabytes downloaded from a third-party server — that is to say, precisely what this site forbids itself. A tool offering them “locally” would fetch the model from a CDN on the first click, and the promise would end there.
Camera RAW files and iPhone HEIC files do not open here: decoding them would require a specialised library. The file is refused when dropped, with its reason, rather than when saving.
What is not guaranteed, and must be said
The spotting of active content displayed when a document is opened is done on the raw bytes. A PDF whose objects are compressed — the most common case — can hide a script from that sweep. This warning warns, it does not protect. The protection is never to execute anything from the document and to rebuild the output, which every export does.